Grant: Discovery Grants Program - Individual

Explainable Artificial Intelligence (XAI) is one of the most effervescent areas of research in Artificial Intelligence (AI) and Machine Learning (ML), which is due to the fact that in general the internal mechanisms of AI and ML systems are difficult to understand and their outputs are difficult to explain. Stakeholders expect some sort of explanation when decisions or results from an AI system affect them. Furthermore, legislation is being introduced, demanding for more Ethical AI systems. This research proposal starts from different approaches the applicant has been developing to confront XAI problems, namely: (a) Applications of causality-related concepts and techniques (e.g. counterfactual interventions, causal responsibility); (b) Introduction and analysis of score-based explanations that assign numbers to feature values in order to reflect their relevance for the outcome at hand, including a new Resp score and the previously existing Shap score. (c) Answer-Set Programming-based specification of counterfactual explanations and scores, and reasoning therewith. Although progress can be exhibited can be exhibited for these different approaches, there are still important open problems and research directions, two of which we briefly describe below.  An important problem is that of of bringing domain knowledge and semantics into the definitions and computations of these explanations and scores. So as building machine learning models may and should benefit from using explicit and general domain knowledge, as opposed to basic data, explanation methods should also take advantage of that kind of additional knowledge. Since domain knowledge is naturally captured by means of logic-based and probabilistic knowledge representation languages, this research goes in the direction of neuro-symbolic AI and ML (in this case, "neuro" standing for any form of algorithmic ML model). Causality is starting to play a prominent role in ML since correlations implicit in the data may not be enough for building accurate, generalizable and robust AI and ML models. Furthermore, causality has been for centuries at the very basis of explanations of phenomena in general. It has been a subject of investigation in AI and ML, but not much has been done in term of applying causality to these areas, in particular, to explain the outcomes of AI and ML systems. On another side, different forms of explanations have been investigated in AI, most prominently, via model-based diagnosis. The connections between causality and other forms of explanation in AI, and also the mathematical and computational properties of causality-based explanations have not been investigated in depth, particularly in XAI. We plan to address these problems in general, and also with special attention in the context of explainable AI. Higher-level causes, incomplete data, and model interpretation will be investigated. We plan to apply the results of our research in business, finance and economics.